AI finance assistant

This project is under NDA, so there is no company name and no industry. The company has many locations and a large database of operations. Managers need figures every day, and there are not enough analysts for all the questions.

How it works

An employee types a question, for example “Where are our write-offs highest?”. The assistant sees what is missing and asks: which period, and in what unit - roubles or a share of revenue. With the answer in hand it builds a database query, checks it, runs it and replies with a table and text. The conversation can go on: “Now in roubles”, “And for June?” - the assistant remembers what it is about.

Security

The language model writes the query but does not run it. Before execution the query is parsed and checked: only reading is allowed, run time is limited and heavy queries are cut off. A set of 23 checks catches attempts to change or delete data. The work runs under a separate database account that can only read.

The whole system sits inside the company’s network and never goes to the internet. Every action - question, query, answer - is logged.

Where the project stands

A working prototype is ready: a browser screen, clarifying questions, conversation memory, query checks and the log. It runs on 18 months of test data. The next step is to connect a model inside the client’s network and run a reference set of questions to measure answer accuracy.

Who it fits

Retail chains, manufacturers, distributors, banks and insurers: companies with a large operational database and a closed network, where managers need answers about the numbers without queueing for an analyst.